Software Bug Number Prediction Based on Complex Network Theory and Panel Data Model

• Published in: IEEE transactions on reliability Vol. 71; no. 1; pp. 162 - 177
• Main Authors: Yang, Shunkun, Gou, Xiaodong, Yang, Minghao, Shao, Qi, Bian, Chong, Jiang, Ming, Qiao, Yongjie
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.03.2022; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Bug number prediction; Codes; complex network; Complex networks; Computer bugs; cross-project (CP) prediction; Data models; Debugging; Linux; Longitudinal studies; Maintenance engineering; Open source software; panel data (PD) model; Percolation theory; Predictive models; Quality assurance; Resource allocation; Software; Software reliability; Software testing; within-project (WP) prediction
• ISSN: 0018-9529; 1558-1721
• DOI: 10.1109/TR.2022.3149658

Accurate software bug number prediction makes software test resource allocation, maintenance, and release time cost efficient. However, it is a challenge to accurately predict the number of software bugs when there fluctuations caused by many uncertain factors faced by the complex software. Considering this, a new method for software bug number prediction based on a panel data model from the perspective of complex networks is proposed in this article. Using complex network theory, we constructed the software code network and calculated the static metrics of the network structure, and the percolation threshold of change in the network structure based on percolation theory as a dynamic metric. These network metrics were then normalized as inputs and a panel data model was used for bug prediction. The proposed method can predict the number of bugs for both within-project and cross-project. Empirical studies were performed on data obtained from 120 releases of eight open-source software projects (Lua, SQLite, Redis, Linux kernel, ant, jmeter, poi, and tomcat), the experimental results indicated that network metrics are effective bug indicators, and the proposed method outperformed ten baseline methods (with an average improvement of 28.05%). This article is expected to provide insights into more smart software quality and reliability assurance.